The first test of AI’s promised independence is proving unexpectedly mundane: making a call, booking an appointment or telling a windowpane from a tabletop. Consumer-agent companies are increasingly routing such jobs to people because even leading models can browse and recommend, yet still stumble when judgment and real-world follow-through are required.¹

At Dazzle AI, Marissa Mayer says the company uses frontier models for suggestions and candidate lists, then checks their work before delivering a service it can “stand behind.”¹ She recalled an agent that mixed up glass for a window with glass for a tabletop and began trying to determine whether the tabletop would fit in the window. A human resolved it within seconds; the agent might have consumed vast computing resources without reaching an answer.¹

That handoff model is spreading. Duckbill has built an API to send likely agent failures to contract workers, while Wajo AI uses people for phone calls and CAPTCHA hurdles. Their argument is practical: collaborative AI keeps tasks moving. But the workaround also sharpens privacy concerns, particularly when users may not realize a human has entered an interaction. Meta has said such human-assisted testing is part of product development, while adding that people do not currently complete Muse users’ requests.¹

The same tension has surfaced in a far more rarefied arena: mathematics. After OpenAI released hundreds of AI-generated results, researchers said they were struggling with sparse writeups, weak context and the laborious task of verification.²

Sam Altman has acknowledged the disruption but insists publication should continue. “Humanity deserves fundamental knowledge, and we should not be sitting on a major fundamental science discovery,” he said, while conceding OpenAI must improve how it communicates results.² Yet the company has also withheld details of a claimed Yang-Mills result on advice from selected experts, frustrating mathematicians who see gatekeeping where OpenAI preaches disclosure.²

From household errands to frontier proofs, the pattern is the same: AI can accelerate the attempt, but people still have to land it safely—and decide when the public can trust the result.¹²